Why the Law Keeps Arriving Too Late for AI Surveillance
In July, I argued in these pages that terms of service can’t protect citizens from AI surveillance, and that the job belongs to law. That raises a harder question. If law is the answer, why does it keep arriving after the cameras are already up, and what would it take to fix that?
AI Has Changed What Surveillance Is
Traditional surveillance was constrained by the cost of attention. A camera could record, but a human had to watch, and human attention is expensive. East Germany’s Stasi had more than 91,000 full-time employees and about 189,000 unofficial collaborators at its height, and even that apparatus couldn’t watch everyone.
AI removes that constraint. Interpretation is now as cheap as collection, so everything captured can be analyzed, cross-referenced, and held against future queries.
Surveillance becomes retrospective. Archived footage and records can be searched after the fact, so a person can be tracked backward through time the moment they become interesting to someone. It becomes inferential. Modern systems claim to infer emotional state or political leanings from seemingly harmless data, like gait, voice, purchasing patterns, or typing cadence. It fuses facial recognition, license plate readers, financial records, and social graphs into a single, queryable picture of a life.
The practical obscurity that protected ordinary people throughout history, the simple fact that nobody was watching, is gone.
Surveillance Arrives Through Procurement
Part of the lag is the pacing problem. Technology iterates in months, legislation takes years, and case law takes decades. Laws written for facial recognition miss gait recognition, voice fingerprinting, and re-identification from supposedly anonymized data, and vendors exploit those gaps deliberately.
The deeper problem is the route these tools take. Surveillance technology typically enters through procurement. A police department buys a tool, and no governing body ever votes on it. By the time lawmakers notice, the capability is operational, constituencies have formed around it, and the argument has shifted from whether to deploy to how much to constrain. Governments are also the customers for these systems, which leaves regulators conflicted as well as outpaced.
The lesson I draw is that rules written around specific technologies age badly, while process requirements endure. Transparency, auditability, authorization, and redress apply whichever technique arrives next. As such, we should regulate the verbs, not the nouns.
Normalized Surveillance Has Lasting Costs
The best-documented cost is the chilling effect. People who believe they’re watched self-censor.
For example, after the 2013 Snowden revelations, Alex Marthews and Catherine Tucker found that Google searches for privacy-sensitive terms declined, and Jon Penney found that traffic to terrorism-related Wikipedia articles fell nearly 30%. Elizabeth Stoycheff also found that awareness of government monitoring made people less willing to voice minority opinions online.
Most reform begins as deviance, and a society that discourages dissent loses the variation that drives progress.
The infrastructure outlives the intentions of those who built it. Systems installed by well-meaning administrations remain available to less scrupulous successors. The relevant question about any surveillance system is whether you’d trust the worst plausible future operator.
The Fix Is Verification at Scale
The largest gap is verification infrastructure. We have principles in abundance. What we lack are mechanisms. My work with the IEEE on transparency standards, including IEEE7001 on the transparency of autonomous systems, starts from the conviction that trust must be engineered into systems instead of appended in policy documents. That means tamper-evident logging of how surveillance systems are queried and by whom, independent audit access, and certification regimes with real consequences.
It also means safeguards most jurisdictions still lack. Purpose limitation with teeth would keep a system procured for finding missing children from becoming a tool for tracking protesters. Retrospective searches need judicial authorization, because querying a year of stored footage about a person is a search and should be treated like one. Mandatory incident reporting and sunset clauses would keep misuse visible and force capabilities to be justified over time.
Where the Line Belongs
Proportionality remains the right starting point. Surveillance should target particular people, based on particularized suspicion, with independent authorization and review. Dragnet collection on entire populations fails that test no matter how it is justified, because it inverts the presumption of innocence.
I’d add that the line is a matter of process as much as it is place. Who authorizes collection, who audits it, how long it persists, and what recourse exists when it goes wrong matter as much as what’s collected.
The security versus privacy framing is also a false choice. Centralized biometric repositories are irresistible targets. The 2015 breach at the U.S. Office of Personnel Management exposed records on 21.5 million people, including roughly 5.6 million fingerprint records. A state that hoards data on its citizens has often just built the weapon and left it lying around.
What the Next Five Years Should Look Like
Realistically, I expect an ecosystem that begins to function, short of a single global framework. The Digital Omnibus has already pushed the high-risk obligations back to December 2027. The next five years will determine whether the EU AI Act becomes a working global baseline or a cautionary tale about unenforced ambition. Its fines for prohibited practices reach €35 million or 7% of worldwide annual turnover, whichever is higher. If those penalties are enforced seriously, the Brussels effect will pull standards upward elsewhere, as GDPR did.
Beyond that, AI needs an audit profession with the independence and liability that made financial accounting trustworthy. It also needs oversight that runs at machine speed, because quarterly committee reviews can’t keep pace with systems making millions of decisions an hour.
Law is still the answer. Today, almost no one who sets surveillance boundaries is accountable for them. The work of the next five years is changing that.
By Nell Watson, AI Ethics and Emerging Technology Faculty at Singularity University
Photo by Miłosz Klinowski on Unsplash






Comment